Lameness Detection as a Service: Application of Machine Learning to an Internet of Cattle

Abstract

Lameness is a big problem in the dairy industry, farmers are not yet able to adequately solve it because of the high initial setup costs and complex equipment in currently available solutions, and as a result, we propose an end-to-end IoT application that leverages advanced machine learning and data analytics techniques to identify lame dairy cattle. As part of a real world trial in Waterford, Ireland, 150 dairy cows were each fitted with a long range pedometer. The mobility data from the sensors attached to the front leg of each cow is aggregated at the fog node to form time series of behavioral activities (e.g. step count, lying time and swaps per hour). These are analyzed in the cloud and lameness anomalies are sent to farmer's mobile device using push notifications. The application and model automatically measure and can gather data continuously such that cows can be monitored daily. This means there is no need for herding the cows, furthermore the clustering technique employed proposes a new approach of having a different model for subsets of animals with similar activity levels as opposed to a one size fits all approach. It also ensures that the custom models dynamically adjust as weather and farm condition change as the application scales. The initial results indicate that we can predict lameness 3 days before it can be visually captured by the farmer with an overall accuracy of 87%. This means that the animal can either be isolated or treated immediately to avoid any further effects of lameness.

SmartHerd management: A microservices‐based fog computing–assisted IoT platform towards data‐driven smart dairy farming

“…The behavioral trend of young and old cows for the three activities is as shown in Figure A and Figure B, respectively. Our initial work has presented age‐based clustering of cows combined with data analytics to detect anomalies in their behavior and microservices‐based application flow for integration of specific services such as lameness in dairy cattle …”

Invited review: Cattle lameness detection with accelerometers

“…Steps per hour alone was reported to be marginally predictive of lameness with an area under the curve of 0.6 by Kamphuis et al (2013). Byabazaire et al (2019) reported lying time, step count, and swaps (changes in behavior) as a basis for a detection model with a classification sensitivity of 89.7% and a specificity of 72.5%. Beer et al (2016) reported greater sensitivity (90.2%) and specificity (91.7%) when using standing bouts and speed, a measure of gait, not behavior.…”

“…Several behavior variables were identified as indicative of lameness. These were activity/walking duration (Thorup et al, 2015), step count (Byabazaire et al, 2019), the ratio of day to night time activity (Van Hertem et al, 2013;Schindhelm et al, 2017), standing bouts and swaps (changes in behavior; de Mol et al, 2013;Beer et al, 2016;Byabazaire et al, 2019). Behavior variables that are not listed here were judged to have evidence that showed them insufficient or too inconsistent (i.e., lying time) to warrant inclusion.…”

“…The variables identified in the previous section are recorded by sensors at different frequencies which can be aggregated and summarized in various ways (400 Hz, 10 Hz, per step, per hour, per day, or per week). Weekly behavior summaries were reported by Garcia et al (2014); daily behavior summaries were reported by Kamphuis et al (2013); Beer et al (2016); King et al (2017), andByabazaire et al (2019); and within day behavior information was reported by Garcia et al (2014) and Grimm et al (2019). Other studies have reported variables specifically from when cows were walking (Thorup et al, 2015), standing (Thorup et al, 2015), and lying (Kokin et al, 2014).…”

“…Other studies have reported variables specifically from when cows were walking (Thorup et al, 2015), standing (Thorup et al, 2015), and lying (Kokin et al, 2014). There is no clear indication as to which of these is the most appropriate, except to observe that the most successful of these behavior-based studies in terms of lameness detection performance used daily summaries (Beer et al, 2016;Byabazaire et al, 2019).…”